AWS agents will suggest your new flights. Code decides what gets booked.
The New Stack Meredith Shubel
AWS wants AI agents to suggest flight fixes, but code makes the final booking. That split matters because it keeps money and changes under a checkable workflow, not an agent’s guess.
Based on reporting by The New Stack, Meredith Shubel — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
AWS has put out a new Step Functions pattern that gives AI agents a useful job in airline rebooking without handing them the keys to the reservation system. In the setup, Amazon Bedrock AgentCore agents can suggest new itineraries and draft compensation messages after a disruption. But the actual reservation changes and any payment stay with deterministic code. AWS’s own line is blunt: agents propose, and deterministic code validates.
That distinction is the whole point. Airline rebooking is tedious, especially when a cancellation means hundreds of passengers need new itineraries. AWS says Step Functions can take on the orchestration work — fan-out, validation, routing, retries — while the agents focus on reasoning about options. The workflow keeps an execution history too, which matters when someone eventually asks why a certain change was made.
The company is also making a broader argument about how agent systems should work. On the same day, it published a case study about Abnormal AI and said agents need a compute environment where they can calculate, process data, and verify work before returning results. In that example, Abnormal AI uses Amazon Bedrock AgentCore Code Interpreter, a serverless runtime for running code dynamically, to support real-time inline email threat detection. AWS is basically saying that an LLM can be smart, but it still needs a scratchpad when the job turns into counting, structured reporting, or verification.
And AWS is not alone in drawing that line. Perplexity’s Portable Computer, which the source says shipped last month, uses deterministic software to decide whether model actions should run at all. The common thread is simple: let models suggest, but don’t let them act unchecked. That is a more useful pattern than pretending reasoning alone is enough for anything with money, accounts, or audit trails attached.
My take — AI-written commentary, not fact-checked reporting
This is the rare sensible AI pitch: keep the model in the brainstorming chair and let code hold the wallet. The industry loves to talk about “agentic” systems as if trust is a vibe; it isn’t. Deterministic checks are the boring part, which is exactly why they’re the part worth betting on.
Read more about this at: The New Stack